Molecular Dynamics Acceleration on Many-core Architectures
TimeTuesday, July 246:30pm - 8:30pm
DescriptionComputationally rigorous molecular dynamics codes must be employed for cutting edge materials investigations. This work describes the enhanced parallelism available on the Intel Xeon Phi in comparison to performance on the faster, but less multithreaded, Broadwell processor. Through specific understanding of architectural nuances, evinced through in-depth exploration of compiler and environment variable options coupled with thorough performance analysis, the tuning of the LAMMPS molecular dynamics application is descriptively presented.
The benchmarks target several different type of molecular dynamics simulations such as: 1) Lennard-Jones liquid, in which relatively simple atomic interactions are computed for a large number of neighboring atoms, 2) embedded-atom-model (EAM) solid, in which more complex potential energy functions are used to approximate the behavior of (typically metallic) atoms, and 3) Stillinger Weber potential, typically used to approximate the behavior of silicon and other tetrahedral phases of matter, and 4) bead chain systems used to model polymeric systems. The different aspects of these models stress the different aspects of hardware performance such as applicability to vectorization, memory bandwidth, and cache management. Analysis and reporting of the benchmark results will enable other researchers to target the optimal hardware and compute capacity for their simulations.
This application of advanced computing infrastructure should dovetail with the purpose of PEARC, demonstrating the creativity made feasible on many-core platforms by empowering state-of-the-art materials simulations.